What You Will Do
Training platform
Work with the client's infrastructure team on the GKE training platform (GPU node pools, job queueing and scheduling, quota and capacity) from the training workload's side.
Diagnose and fix platform failures at the root (for example a controller that keeps running out of memory) rather than restarting them.
Tune multi-node GPU networking (RDMA/GPUDirect, NCCL) for training throughput.
Distributed training
Run, scale and debug multi-node distributed training with PyTorch (FSDP, DeepSpeed, Megatron or NeMo-style) and Ray Train.
Profile jobs and remove bottlenecks across compute, memory, input pipelines and interconnect.
Benchmarking and evals
Design and run performance benchmarks for the cluster and training stack: throughput, scaling efficiency, cost per run.
Build and maintain model evaluation harnesses so changes to the platform or training recipes are measured, not assumed.
Instrument training and the platform with telemetry and monitoring, and keep end-to-end pipeline tests running against what users actually run.
Ownership on the client team
Treat the platform as your own: proactively find issues, plug gaps and propose solutions before anyone asks, keep the team informed of status and risks, and look after the scientists who use it.
Explain and defend your design decisions to client engineers and scientists in detail.
- Stay effective while the client's priorities and team structure change.
- Your Technical Toolkit:
- Core Languages: Mastery of Python and shell scripting
AI/LLM Ecosystem: Extensive experience with Google Gemini, GPT-4, or LLaMA; deep knowledge of Prompt Engineering and Fine-tuning
Data & Search: Expertise in Vector Databases (Vertex AI Vector Search, pgvector, etc.) and implementing Semantic Search
Infrastructure: Hands-on experience with Google Cloud (Vertex AI) and building scalable software architectures
Frameworks: Proficiency in LangChain, LlamaIndex, or similar orchestration layers
- Mindset: A strong software engineering foundation—you write clean, maintainable code and understand the full SDLC
- Basic Qualifications:
- 5+ years of engineering experience, with recent hands-on work on GPU training infrastructure and large-scale training.
Strong Python software engineering: production-quality code, tests and code review. Coding is central to this role, not an afterthought.
GKE and GPU scheduling: Kubernetes on Google Cloud with GPU node pools, job queueing (Kueue or similar), and capacity and quota management.
Distributed PyTorch and Ray Train: has personally run and debugged multi-node training jobs.
RDMA and GPU networking: GPUDirect/RDMA, NCCL tuning, multi-node GB200 (or H200) topology.
Benchmarking and evals: has designed performance benchmarks and model evaluations, and used them to make decisions.
Deep enough understanding to explain how and why a system works, without leaning on AI coding tools to carry the reasoning.
- A track record of visible ownership: present, communicative, and accountable on a demanding team.
- Personal Attributes:
- Nice to have
Healthcare or life-sciences data (genomics, clinical, biomedical), and working under HIPAA.
Observability and telemetry for training platforms: metrics, tracing, alerting, end-to-end pipeline testing.
